[clang] [llvm] [clang][ScanDeps] Allow PCHs to have different VFS overlays (PR #82294)

[Michael Spencer via cfe-commits]

================
@@ -65,11 +66,25 @@ static void optimizeHeaderSearchOpts(HeaderSearchOptions &Opts,
     llvm::DenseSet<const serialization::ModuleFile *> Visited;
     std::function<void(const serialization::ModuleFile *)> VisitMF =
         [&](const serialization::ModuleFile *MF) {
-          VFSUsage |= MF->VFSUsage;
           Visited.insert(MF);
-          for (const serialization::ModuleFile *Import : MF->Imports)
-            if (!Visited.contains(Import))
-              VisitMF(Import);
+          if (MF->Kind == serialization::MK_ImplicitModule) {
+            VFSUsage |= MF->VFSUsage;
+            // We only need to recurse into implicit modules. Other module types
+            // will have the correct set of VFSs for anything they depend on.
+            for (const serialization::ModuleFile *Import : MF->Imports)
+              if (!Visited.contains(Import))
+                VisitMF(Import);
+          } else {
+            // This is not an implicitly built module, so it may have different
+            // VFS options. Fall back to a string comparison instead.
+            auto VFSMap = PrebuiltModuleVFSMap.find(MF->FileName);
+            if (VFSMap == PrebuiltModuleVFSMap.end())
+              return;
----------------
Bigcheese wrote:

I think it's also possible that you used an explicit module in the original driver command, or prebuilt module path. The idea of silently ignoring unknown modules here was to preserve the existing behavior of ignoring them. I would eventually like to make this an error, but doing this works in most cases now, so I don't want to do that yet. It would probably be good to make it a warning in this patch though, I don't expect this return to fire unless something weird is going on.

https://github.com/llvm/llvm-project/pull/82294